Research And Application Of Automatic Portrait Matting Algorithm Based On Deep Learning

Portrait matting is a technology that identifies the portrait in images,including head,bust and whole body,and accurately predicts the opacity in the junction of portrait and background.Most of portrait matting tools complete the matting process by way of providing color sample information by drawing a trimap or strokes image based on user interaction,which is tedious and time-consuming.For professionals not having matting knowledge,a satisfactory matte requires them to draw many trimaps or manual maps,and most of the portrait matting tools are incapable of processing images with a small contrast ration between foreground and background colors.While the portrait segmentation without user interaction is a rough and forced segmentation of portraits,which results in no smooth transaction between the portrait and background as well as not very accurate segmentation of foreground(hairs)with a complex structure.This paper summarized the latest research results of image matting and deep learning at home and abroad,focused on the utilization of image matting algorithm and per-pixel task algorithm based on deep learning,and analyzed the problems in per-pixel task algorithm.The automatic portrait matting algorithm based on deep learning was proposed in view of problems such as low effectiveness,requiring manually annotated trimap and the difficulty in improving the accuracy caused by relying on color as the main basis for most of image matting.The main work accomplished in this paper was as follows:(1)This paper referred image matting DAPM and DIM algorithms to propose a fast automatic portrait matting algorithm based on the multitasking convolution nerve.In addition,taking the public database as the test set,comparison was carried out between Deeplab and IMFM and DAPM algorithms.The results showed that the former was superior to the comparing algorithms in terms of the evaluation indexes of running time,SAD,MSE and Gradient.(2)Taking the typical Deeplabv3 network with end-to-end per-pixel task algorithm as the research object,this paper summarized the advantages and disadvantages of its used dilated convolution and separable convolution as well as their influence on network performance.LASPP with a lighter weight and a wider receptive field was proposed after analysis on the restricted receptive field and " checker-board " issue due to "insufficient depth" of ASPP;the advantages of SE module were utilized to combine with the separable convolution.Based on the experimental results in PASCAL VOC2012,a public image segmentation dataset,it was verified that LASPP has the characteristics of less time consuming and more convolution integration as well as more abundant details and strong robustness in comparison with ASPP;SE module could be used with the separable convolution model to effectively promote the network performance.(3)An automatic portrait matting algorithm was designed,which was composed of the trimap predication network and the Alpha image predication network.The proposed trimap predication network utilized its advantage of multiple network output to improve the evaluation indexes of PA and m IOU by 0.029 and 0.103,respectively in comparison with the trimap predicted single network output.The quadratic differential difference loss function of Alpha image was added on the basis of the loss function in DIM algorithm,with the public data as the validation set,and the evaluation indexes of SAD and MSE were decreased by 2.13 and 0.002,respectively.(4)Finally,based on the automatic portrait matting algorithm proposed in this paper,the identification photo generation system was developed,which was applied in self-service photography halls for identification photo and the identification photo generation software based on photos.